Selection schemes between neural-network (NN) and multiple-regression (MR) outputs of a virtual metrology system (VMS) are studied in this paper. Both NN and MR are applicable algorithms for implementing virtual-metrology (VM) conjecture models. A MR algorithm may achieve better accuracy only with a stable process, whereas a NN algorithm may have superior accuracy when equipment property drift or shift occurs. To take advantage of both MR and NN algorithms, the simple-selection scheme (SS-scheme) is first proposed to enhance the VM conjecture accuracy. This SS-scheme simply selects either NN or MR output according to the smaller Mahalanobis distance between the input process data set and the NN/MR-group historical process data sets. Furthermore, a weighted-selection scheme (WS-scheme), which computes the VM output with a weighted sum of NN and MR results, is also developed. This WS-scheme generates a well-behaved system with continuity between the NN and MR outputs. Both the CVD and photo processes of a fifth-generation TFT-LCD factory are adopted in this paper to test and compare the conjecture accuracy among the solo-NN, solo-MR, SS-scheme, and WS-scheme algorithms. One-hidden-layered back-propagation neural network (BPNN-I) is applied to establish the NN conjecture model. Test results show that the conjecture accuracy of the WS-scheme is the best among those solo-NN, solo-MR, SS-scheme, and WS-scheme algorithms.

Over the past few years, the virtual metrology (VM) technology has been proposed and developed and several VM related papers have been published. These proposed VM architectures are mainly applied for conjecturing single-stage or direct processing quality. However, the processing quality of some of manufacturing processes cannot be measured directly. Those single-stage VM architectures may not be applied to handle this indirect VM problem accurately. To properly handle this indirect VM problem, the paper proposes a dual-stage indirect VM architecture (DIVMA), which involves two process tools, to handle the indirect VM problem. In Stage I, the direct VM model of the fore-tool is established as usual, and then the indirect VM model of the rear-tool is built in Stage II that involves the Stage-I VM output. The concept of virtual cassette is also proposed in the paper. The chemical vapor deposition (CVD) process of a fifth generation TFT-LCD fab in Chi Mei Optoelectronics Corporation (CMO), Taiwan is adopted to test and compare the conjecture accuracy without and with the virtual cassette and the proposed DIVMA. Experimental results demonstrate that the virtual-cassette concept is valid for improving VM conjecturing accuracy; also this DIVMA is adequate for handling the indirect VM problem that has a coupling effect between those two subsequent tools.

This work proposes an advanced key-variable selecting method, the neural-network-based stepwise selection (NN-based SS) method, which can enhance the conjecture accuracy of the NN-based virtual metrology (VM) algorithms. Multi-regression-based (MR-based) SS method is widely applied in dealing with key-variable selecting problems despite that it may not guarantee finding the best model based on its selected variables. However, the variables selected by MR-based SS may be adopted as the initial set of variables for the proposed NN-based SS to reduce the SS process time. The backward elimination and forward selection procedures of the proposed NN-based SS are both performed by the designated NN algorithm used for VM conjecturing. Therefore, the key variables selected by NN-based SS will be more suitable for the said NN-based VM algorithm as far as conjecture accuracy is concerned. The etching process of semiconductor manufacturing is used as the illustrative example to test and verify the VM conjecture accuracy. One-hidden-layered back-propagation neural networks (BPNN-I) are adopted for establishing the NN models used in the NN-based SS method and the VMconjecture models. Test results show that the NN model created by the selected variables of NN-based SS can achieve better conjecture accuracy than that of MR-based SS. Simple recurrent neural networks (SRNN) are also tested and proved to be able to achieve similar results as those of BPNN-I.

For maintaining high stability and production yield of production equipment in a semiconductor fab, on-line quality monitoring of wafers is required. In current practice, physical metrology is performed only on monitoring wafers that are periodically added in production equipment for processing with production wafers. Hence, equipment performance drift happening in-between the scheduled monitoring cannot be detected promptly. This may cause defects of production wafers and the production cost. In this paper, a novel virtual metrology scheme (VMS) that is based on a radial basis function neural network (RBFN) is proposed for overcoming this problem. The VMS is capable of predicting quality of production wafers using real-time sensor data from production equipment. Consequently, equipment performance abnormality or drift can be detected timely. Finally, the effectiveness of the proposed VMS is validated by tests on chemical vapor deposition (CVD) processes in practical semiconductor manufacturing. It is therefore proved that RBFN can be effectively used to construct prediction models for CVD processes